Structure Prompts Around Sources, Done, and Evidence

Most prompts spend their first lines on framing: a role, a reminder that the task matters, a request to be careful. Those lines rarely change the output.

The Claude Code best-practices guidance points at three things that do change the output. Put your words there instead.

1. Point to sources

Tell Claude where the answer lives: a folder, a file, a ticket, a log, a previous PR.

The client notes are in onboarding/acme/. Use those, not the old template.

Claude can find things on its own, but every search costs time and context. A direct pointer skips that.

2. Define done

Describe what the finished output looks like, concretely enough that Claude can tell when it has arrived.

Done means a one-page proposal covering their three pain points,
with the price at the bottom.

Without this, Claude stops when the result looks plausible.

3. Ask for evidence, not assertions

Give Claude a check it can run, and ask it to show the result instead of claiming success: test output, the command it ran, a screenshot.

Before you finish, check every number against the notes
and flag anything you cannot back up.

For code, this is the verification loop: tests, a build, a typecheck, a screenshot comparison.

Before and after

Before:

You are a senior copywriter with 20 years of experience.
This is very important. Write the client proposal.

After:

Write the client proposal. Notes are in onboarding/acme/.
Done is a one-page proposal covering their three pain points with price at the bottom.
Before you finish, check every number against the notes and flag anything you cannot back up.

Roughly the same length. Every line in the second version changes what Claude does.

Where this comes from

These three points come from the Claude Code best-practices docs: point to sources, give Claude a check it can run, and have it show evidence rather than asserting success. Persona lines are not forbidden, they are just not where the leverage is.